AI Foundry Service (AFS) of Taiwan AI Cloud is a comprehensive enterprise-level generative AI solution, providing an end-to-end service from large language model optimization to inference cloud/on-premise deployment. This includes the “AFS Platform+AFS Cloud” for helping enterprises fine-tune proprietary models and establish inference services, and the “AFS ModelSpace” which offers various Traditional Chinese enhanced Formosa Foundation Models (FFM) and reliable open-source models for quick cloud inference. All these services can be operated using a No-code platform.

If users need to establish proprietary models, please quickly select FFM base models and high-performance throughput computing configurations through the “AFS Platform” and complete the fine-tuning training of hundreds of millions of tokens within a few hours. Simultaneously, the “AFS Cloud” automatically creates inference service endpoints and independent API URLs, allowing for enterprise generative AI application integration and Chat/Playground testing.

For enterprises that do not need model fine-tuning, please choose the latest FFM or open-source models on “AFS ModelSpace” and directly call APIs or perform inference tasks via Chat/Playground. “AFS ModelSpace” also offers different billing modes, by the number of tokens used (by tokens) or by model usage in GPU hours.

1. AFS Platform+AFS Cloud: Provides a billing mode based on “model building time”, with costs varying according to the selected base model.

2. AFS ModelSpace: Provides dual billing modes based on “number of tokens used” and “model building time”, with costs varying according to the selected base model.

For more information, please refer to the AFS Pricing.

1. For users who need to establish proprietary large models, please choose “AFS Platform+AFS Cloud” to help fine-tune proprietary models and establish inference services.

2. For users without model fine-tuning needs, please choose “AFS ModelSpace” to interact with various FFMs and open-source large models in real-time.

3. For users with sensitive data and security compliance needs, please choose to deploy proprietary large models on-premises in enterprise data centers/private clouds with the “FFM Licensing”.

For more information, please refer to the AFS Services.

The AI Foundry Service (AFS) of Taiwan AI Cloud is tailored for enterprise users, meeting their security, SLA, maintenance, and compliance needs. AFS offers a full range of services, including POC verification, customized model optimization (Fine-tuning), hybrid on-premise and cloud model deployment, and inference. Users can choose services based on actual needs. Rest assured, Taiwan AI Cloud does not retain any customer data or records from using AFS services or any product interactions, nor does it use them as training data for FFM.

Additionally, all AFS services are built under national information security regulations and hosted in non-government A-grade computing facilities, ensuring that user services and data meet the highest security standards. For higher sensitivity and regulatory requirements, users can choose Taiwan AI Cloud’s “FFM Licensing” to deploy proprietary models on-premises, ensuring no internal data leakage.

If the specific internal and external data volume in a particular field for an enterprise exceeds 1 billion tokens, it is best to use a two-stage optimization training with Pretrain-SFT to establish the best enterprise-specific domain model. Adding a new dictionary that includes enterprise-specific domain knowledge, such as patent translations and specialized lexicons, ensures that certain terms correspond to specific translations or domain-specific meanings. Building a specialized dictionary during the pre-training stage effectively integrates this new dictionary into the model.

If insufficient data is available for pre-training, SFT can still be conducted directly. Still, the SFT requires at least 270,000 tokens (for partial parameter fine-tuning) to 570,000 tokens (for full parameter fine-tuning).

 The QA pair data set should be in jsonl format, with each data entry having a total token count not exceeding 4096.

The Formosa Foundation Model (FFM) is a Traditional Chinese enhanced large language model (LLM) developed by Taiwan AI Cloud based on LLM technology. The technical R&D team of Taiwan AI Cloud, leveraging years of NLP experience, has enhanced the performance of pre-trained models using LLM technology. It demonstrates high-quality understanding and performance in Traditional Chinese semantics and knowledge domains while maintaining the capabilities of the native model. Tailored to meet the diverse application needs of Taiwanese enterprises in different fields, FFM provides high-quality text generation results and possesses comprehensive world knowledge and multilingual capabilities.

The FFM model series offers Traditional Chinese enhanced versions of various open-source models, from BLOOM, Llama2, and Mistral to the latest Llama3. In terms of model specifications, it provides the most complete parameter configurations for each model, including BLOOM-176B, Mixtral-8x7B, Llama3-70B, and more. The Phi-3 version of FFM is also under development, so stay tuned.

In the rapidly evolving LLM application needs, Taiwan AI Cloud is committed to providing the latest, most complete, diverse, and high-quality large language model specifications and application services to accelerate enterprise AI 2.0 implementation.

Yes, it supports. Through the customization method provided by LangChain (please refer to the following link: https://python.langchain.com/en/latest/modules/models/llms/examples/custom_llm.html), application system developers can write custom LLM wrappers according to the manual and specify the location of the deployed FFM.

Welcome you to register as a Taiwan AI Cloud member to enable the “Experience Project Application” and receive 10,000 credits for 60 days of free use, quickly experiencing “AFS ModelSpace” and other cloud platform products. For more details, please refer to the free trial project description.

Create an account to enjoy 60 days of experience for 10,000 credits for free now! Sign up here.

If you need further assistance, please contact us via sales@twsc.io or https://tws.twcc.ai/contact-us/

OpenAI’s GPTs and its GPT marketplace/store primarily target individual users, aiming to create personal GPT services. In contrast, Taiwan AI Cloud mainly provides enterprise-level users with reliable and controllable open-source large language models and generative AI solutions, while also offering on-premises inference services that meet enterprise compliance, security, and audit requirements.

The main differences are as follows:
1. Nocode Platform – Taiwan AI Cloud offers no-code model training, fine-tuning, and inference services.
AFS is a complete no-code platform, allowing an easy and quick start from model training to inference deployment since NVIDIA provides the NeMo development framework and DGX Cloud, which requires writing jobs and related programs.

2. Trusted Open-Source Large Language Models – Taiwan AI Cloud provides trusted open-source large language model solutions with local Taiwanese knowledge and enhanced Traditional Chinese corpus.
FFM is a large language model trained with local Taiwanese knowledge and enhanced Traditional Chinese corpus since NVIDIA’s models are pre-trained models mainly supporting English tasks without enhancement for Taiwanese knowledge and Traditional Chinese corpus.

3. Pricing Model – Taiwan AI Cloud provides different pricing models for model fine-tuning and inference service deployment based on the selected services.
Using NVIDIA AI Foundry Service requires paying for the NVIDIA AI Enterprise Software License and cloud deployment infrastructure costs. Additionally, using DGX Cloud incurs high monthly usage fees.